gpt/utils

Utilities

This directory contains utility modules and helper scripts for the GPT implementation.

Contents

• sparse_attention.py: Implementation of sparse attention mechanisms for billion-token context support
• dynamic_context.py: Dynamic context management with middle token eviction for unlimited context length
• version_manager.py: Utilities for managing different GPT versions and configurations
• setup_branches.sh: Script to set up the branch structure for different GPT versions

Usage

Version Manager

from utils.version_manager import create_model, enable_unlimited_context

# Create a specific GPT version
model = create_model('gpt4')

# Enable unlimited context
model = enable_unlimited_context(model, preserve_begin_ratio=0.1, preserve_end_ratio=0.2)

Dynamic Context Management

from utils.dynamic_context import DynamicContextManager, DynamicContextConfig

# Create a custom configuration
config = DynamicContextConfig(
    max_memory_mb=8192,
    begin_tokens_to_preserve=1000,
    end_tokens_to_preserve=2000
)

# Initialize the context manager
context_manager = DynamicContextManager(config)

Branch Setup

To set up the branch structure for different GPT versions:

cd utils
./setup_branches.sh